ImmersedPrivacy - Social Context Rating Error (Audio as Text): leaderboard
Social-context privacy norms in ImmersedPrivacy (Wang et al., 2026), Tier 2 rating mode: the agent sees egocentric images of a simulated room, gets a text description of the scene's ambient sound instead of the audio, and rates each candidate action's appropriateness for the current social state on a 1-5 scale. Mean absolute error against the mean human rating (0-4); lower is better.
Metric: Mean absolute error (Tier 2: the model's 1-5 appropriateness rating of each candidate action against the mean human rating; audio replaced by a text description of the soundscape (Tier 2) or the dialogue transcript (Tier 3); lower is better). Source: arxiv.org. Saturation forecast: Around 2029. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (2026-03-05) | 1.03 | #84 |
| 2 | Gemini 3 Flash (Preview) (Low) | 1.24 | #141 (Gemini 3 Flash (Preview)) |
| 3 | Gemini 3 Flash (Preview) (High) | 1.25 | #141 (Gemini 3 Flash (Preview)) |
| 4 | Qwen 3.5 27B (Thinking) | 1.26 | #358 (Qwen 3.5 27B) |
| 5 | GPT-4o Mini (2024-07-18) | 1.32 | #1512 |
| 6 | Gemini 3.1 Pro (Preview) (Low) | 1.37 | #91 (Gemini 3.1 Pro (Preview)) |
| 7 | Qwen 3.5 27B (Non-reasoning) | 1.39 | #358 (Qwen 3.5 27B) |
| 8 | Gemini 3.1 Pro (Preview) (High) | 1.44 | #91 (Gemini 3.1 Pro (Preview)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=immersedprivacy-social-context-rating-error-audio-as-text · How It Works · Data refreshed daily, snapshot 2026-11-12.